[LLVMdev] LLVM analysis of Linux programs

[Rafael Ávila de Espíndola]

On 11-06-16 4:20 PM, Ben Farley wrote:
> Hey all,
>
> I'm trying to do computations on the LLVM bitcode representations of
> various Linux programs, and I'm running into some troubles. I have the
> source code for various programs (tcpdump, gzip, curl, etc) and I want
> to compile each program down into a single LLVM bitcode file. I've tried
> a few different methods so far. Originally, I attempted to use the
> Makefiles of the projects with llvm-gcc as the compiler (make
> CC="llvm-gcc -emit-llvm"), then manually link all the resulting bitcode
> files into a single one. This works on a few examples, but not all of
> them. Next, I was pointed to Ben Hardekopf's work on pointer analysis;
> it turns out he has a nifty perl script for doing exactly the sort of
> thing I'm trying to do. Unfortunately, that script doesn't seem to work
> on all programs either. It worked quite well on some of my examples, but
> not at all on others.
>
> So basically, I'm wondering if you all have any advice on how to go
> about doing this. Is there some universal method of compiling entire
> source packages into a single LLVM bitcode file, or possibly some mix of
> tools/techniques that will work for the majority of programs? Or is my
> best bet just to hack together the sort of things I've already been trying?


What I would suggest is first setting up a regular LTO build with gold:
http://llvm.org/docs/GoldPlugin.html

Once you have that working, you can manually get the bitcode for the 
program you want by passing -plugin-arg=also-emit-llvm to the linker. 
Something like

$ rm bzip2
$ make bzip2
<will print the like used for linking>
$ <link line> -Wl,-plugin-arg=also-emit-llvm

> Thanks,
> Ben

Cheers,
Rafael